End-to-end walkthrough · Quadrazene™
Document AI — Order Entry Walkthrough
A guided tour of a live Quadrazene session: a customer purchase order arrives by email, the AI extracts and validates it, a human reviews and approves, and a SAP sales order is created — with a full audit trail on every step.
The Mailbox
Inbound PO emails staged for Document AI processing.
Quadrazene polls a monitored IMAP inbox and stages each inbound email. Status badges — NEW, PROCESSING, PROCESSED, FAILED — track every order through its lifecycle. When a processor clicks PROCESS, the chain fires immediately, picks up the email and its PDF attachment, and begins the pipeline.
Mailbox — NEW, PROCESSED, and FAILED orders across multiple customers.
The chain selector shows "order-entry" — the compiled, 7-step workflow. Fetching staged one email, skipped zero duplicates. Every run gets a transaction ID (txn:…) that traces every downstream event.
PROCESSING… — chain fires on the Nordhavn Logistics purchase order.
From this point the pipeline is fully autonomous until it hits a HITL checkpoint. No human touches the data between here and the intake review.
REACTION AWAITING-HITL — the chain pauses for human review.
After the AI extracts the order, the pipeline deliberately stops and waits for a reviewer to verify the data before writing to SAP. This is governance by design, not a failure mode.
AI Extraction
The Insights engine reads any PDF layout — no templates.
atomic-intake-extract calls the language model with the purchase order PDF and a structured extraction prompt. Every field is returned with a confidence score and a citation back to the exact location in the source document. If the model is uncertain about a field, it flags it — the HITL gate catches it.
Workflow shape — intake-pdf-extract → message-writer, two atoms in sequence.
The flowchart shows exactly how the action runs internally. Each atom is auditable, reproducible, and hash-chained into Records. Duration: 16,566 ms for a 2-page PDF (3,039 tokens).
LLM TOKENS: 2,281 — visible in the admin spend panel alongside the action.
Every token consumed is tracked per reaction. The LLM Spend dashboard rolls this up by engine and date so you always know the exact AI cost — no surprise cloud bills.
atomic-intake-extract COMPLETED — 11,118 ms, 2,281 tokens, SUCCEEDED.
The Messages Produced section shows an INTAKEDRAFT record: "PDF extraction complete." This structured draft moves to the next chain step — the HITL intake review.
Human-in-the-Loop
HITL — AI pauses, human decides. Every time.
The chain-root reaction holds the entire workflow while the reviewer acts. Confidence thresholds determine what routes to review automatically — operators tune them per customer and document type. High-confidence, clean orders can be configured to post straight through.
chain-root AWAITING-HITL — HITL REQUEST: intake-pdf-review PENDING.
The chain-root shows 3 of 7 reactions completed. The workflow does not advance until the reviewer acts. The full Reaction Output is available in Tree or JSON for debugging or audit.
Intake Review — TechCorp
Source PDF beside extracted fields — side by side, in one view.
The Inbox Approvals tab surfaces the intake review card. On the left: the original PDF, rendered inline. On the right: every extracted field, colour-coded by confidence. HIGH (cyan) means the field was validated against live SAP master data. The reviewer never leaves the application.
Sales Order Header — Customer, PO number, currency, delivery date. All HIGH.
TechCorp Manufacturing Inc. · PO 7745-Q2-PO-1148 · USD · delivery 2026-06-21. HIGH confidence on all header fields means each was validated against SAP customer master and confirmed clean.
Ship-To and SAP Routing — address, sales org, distribution channel.
Ship-to is matched to SAP partner functions. SAP Routing fields (VKORG, VTWEG, SPART, AUART) are inferred from customer master and shown LOW — they may need override before posting.
Line items and Outbound Email Review — both awaiting action.
Three line items extracted. The reviewer can approve the SAP write and the outbound email as two separate HITL decisions.
Intake Review — Nordhavn Logistics
Six-line EUR order — price correction with documented rationale.
A second purchase order from Nordhavn Logistics GmbH demonstrates the correction workflow. The AI extracted a unit price that differs from the contracted rate. The reviewer edits the field inline, and Quadrazene immediately requires an Override Rationale — capturing who changed what and why into the tamper-evident audit chain.
Nordhavn Logistics GmbH — NL-2026-0617 · EUR · delivery 2026-06-28. All header fields HIGH.
Source PDF from a German procurement system. The AI reads it without any customer-specific template — fully layout-agnostic extraction.
Six cleaning-product line items + Extractor Warnings.
EXTRACTOR WARNINGS flag payment terms ("Netto 60 Tage") and a pallet spec reference that could not be mapped to SAP fields. These are surfaced for the reviewer — not silently dropped.
Price field in edit — reviewer corrects unit price against contract.
The field is editable inline. No separate form, no screen switch. As soon as a value is changed, the Override Rationale box appears.
Override Rationale required — deviation captured into the audit chain.
"Price adjustment applied for BL-CL-001 Allzweckreiniger Citrus and BL-WS-002 Weichspüle items." Text, before/after values, reviewer identity, and timestamp are all hash-chained into Records.
Outbound Email Review — AI-drafted confirmation, second HITL gate.
After the SAP write, Quadrazene drafts the customer acknowledgement. A reviewer approves it before it is sent — the email is never dispatched without explicit human sign-off.
SAP Write
atomic-sap-so-write — the order is real in SAP within milliseconds.
Once approved, the Actions engine POSTs the validated order to SAP via standard OData API. A dedicated technical user writes through a released interface — no GUI scraping, no named-user impersonation. SAP responds with the document number, which is linked back to the source email.
"SAP Sales Order 500000017 created — 6 lines" — Nordhavn order, 195 ms, SUCCEEDED.
PATH: /sap/opu/odata/sap/SADL_V_SALESORDER_BO_CDS/sepm_cds_sales_order — a standard, licensed, sanctioned OData write.
"SAP Sales Order 500000016 created — 3 lines" — TechCorp order, same pipeline.
Two different customers, two different PDF formats, one governed pipeline. The chain, atoms, and audit trail are identical for both runs.
Chain Complete
Customer confirmation sent. Chain closes at 7/7 steps.
The outbound draft uses a deterministic template atom — no LLM call needed. It composes the acknowledgement from the SAP order number. After the reviewer approves it in the Inbox, it is sent. Both HITL gates are recorded with timestamps in the chain timeline.
atomic-outbound-draft — "Re: Order processing" · 46 ms · 0 LLM tokens.
outbound-email-drafter fills a template; message-writer packages it. No AI call needed — the structured order data already exists from the intake step.
order-entry v1.0.0 COMPLETED — 7/7 steps — 2 HITL APPROVED.
Pull email → Extract PDF → HITL: intake APPROVED → Create SAP order → Draft outbound → HITL: email APPROVED → Send email. 408 seconds end-to-end. Every step traceable.
Governance
Every Reactor query fans out to parallel atoms — all auditable.
The Governance engine runs a policy audit alongside every Actions reaction. The actions-sap-operation workflow shows how a natural-language question becomes a parallel fan-out of OData atoms — title-generator and nl-to-sap-action run in parallel, then sap-executor, then action-interpreter and followup-generator. Every atom is hash-chained into Records.
actions-sap-operation — PARALLEL → SEQUENTIAL → PARALLEL workflow.
"How has the distribution of order statuses changed over time?" — the NL question is decomposed into parallel OData atoms in 11,641 ms using 7,768 tokens. The governance-policy-audit reaction fires automatically on completion.
The Reactor
Ask the Reactor. Get live SAP data, AI summary, and auto-generated charts.
The Reactor is a conversational interface to your SAP data powered by the Actions engine. Type a natural-language question — the Reactor translates it to an OData query, executes it against live SAP, and returns structured results with an AI-generated summary and follow-up suggestions. No pre-built dashboard. No static queries.
READ · SALES ORDER ITEMS — 3 live line items + AI summary for order 500000016.
The OData path, filter, and format parameters are fully visible. Results export as CSV or XLSX. The ACTION SUMMARY: "totaling $22,698 gross amount… scheduled for delivery on June 7, 2026."
READ · SALES ORDERS — 16 live orders across TechCorp, Nordhavn, ACME, and more.
Retrieved in seconds from SAP ECC. Orders span multiple customers and currencies. The Reactor auto-routes to the correct SAP system based on the question context.
ACTION SUMMARY — insights, anomalies, and follow-up questions auto-generated.
The summary flags: four recent orders with zero amounts (data quality concern), one order on hold for credit clearance, six currencies, one order with partial billing. Follow-ups suggested automatically.
Charts generated from the same data pull — no configuration required.
Lifecycle status distribution, gross amount by buyer. The Reactor picks the best chart type for the data shape. Charts can be pinned to storylines or exported to PDF.
Cumulative gross over time · order volume by buyer and currency — treemap.
The treemap shows order concentration by buyer and currency at a glance. The area chart shows cumulative revenue over time. The combo chart shows order count vs. average amount by lifecycle stage.
37% EUR · 31% USD · 19% GBP — currency distribution across the portfolio.
Revenue Components by Buyer stacked bar, Sales Order Creation Trend line, and Net Amount vs Tax scatter — all from one natural-language query to a live SAP ECC system.
LLM Cost
Full token-level cost visibility — no black box.
Every LLM call is tracked, every token counted, every dollar attributed. The LLM Spend dashboard rolls up cost by engine, scope tier, and date range. You know exactly what AI costs — at the reaction level — before you get a cloud bill.
$1.13 total · 94,257 tokens · 16 runs — 30-day view.
That covers every Reactor query, every intake extraction, every outbound draft in this demo session. The Jun 7 spike is the Nordhavn order run (16,566 ms extraction on a 2-page PDF). Cost estimated at 1:3 input:output token ratio from current vendor list prices.